Improving Pronoun Translation by Modeling Coreference Uncertainty

Proceedings of the First Conference on Machine Translation, Volume 1: Research Papers, pages 12–20, Berlin, Germany, August 11-12, 2016. c⃝2016 Association for Computational Linguistics Improving Pronoun Translation by Modeling Coreference Uncertainty Ngoc Quang Luong and Andrei Popescu-Belis Idiap Research Institute Rue Marconi 19, CP 592 1920 Martigny, Switzerland {nluong,apbelis}@idiap.ch Abstract Information about the antecedents of pro- nouns is considered essential to solve cer- tain translation divergencies, such as those concerning the English pronoun it when translated into gendered languages, e.g. for French into il, elle, or several other options. However, no machine translation system using anaphora resolution has so far been able to outperform a phrase-based statistical MT baseline. We address here one of the reasons for this failure: the im- perfection of automatic anaphora resolu- tion algorithms. Using parallel data, we learn probabilistic correlations between target-side pronouns and the gender and number features of their (uncertain) an- tecedents, as hypothesized by the Stan- ford Coreference Resolution system on the source side. We embody these correlations into a secondary translation model, which we invoke upon decoding with the Moses statistical phrase-based MT system. This solution outperforms a deterministic pro- noun post-editing system, as well as a sta- tistical MT baseline, on automatic and hu- man evaluation metrics. 1 Introduction Pronoun translation remains a challenge for ma- chine translation (MT), likely because solving cer- tain translation divergencies between source and target pronouns requires non-local information, possibly from one or more sentences before the one that is being translated. In this paper, we fo- cus on the divergencies that occur when translat- ing the English neutral pronouns it and they into French. Depending on their functions (referen- tial or pleonastic) and on their actual antecedents, Source: My cat brought home a mouse that he hunted, and it1 was not dead but it2 was mor- tally wounded. What is the best way to kill it3 humanely? MT: Mon chat a ramen´e `a la maison une souris qui il a chass´e, et il1 ´etait pas mort, mais il2 a ´et´e mortellement bless´e. Quelle est la meilleure fac¸on de le3 tuer humainement? Figure 1: Wrong translations of it into French (1– 3) resulting in a serious misunderstanding. there are almost twenty different lexical items that can serve as translations into French, e.g. for it: il, elle, ce/c’, cela, c¸a, on, le, and others. For instance, in an example from an online dis- cussion forum shown in Figure 1, two referents are mentioned, a cat and a mouse, which are translated in French by nouns with different genders: mas- culine for cat (le chat) vs. feminine for mouse (la souris). The three instances of it, referring to the mouse, should be translated into feminine French pronouns: respectively elle, elle and la (the latter is an object pronoun). However, the online MT system to which we submitted this example trans- lated all of them with the masculine forms, mak- ing the readers think that the author intends to kill his/her cat. The designers of MT systems have …

Content truncated to 500 words